Do I need to regressout uninteresting sources of variation bebofe pseudobul DESeq2 differential expression analysis in single cell data

Question

[alvaroperona1]

Here is the corrected and polished version of your updated question:

I have combined a series of datasets containing a specific cell type. These datasets come from different (but related) tumors, studies, and timepoints. I want to perform a differential expression analysis on these cells based on whether the patient survived or not. The problem is that, during an initial exploration with PCA, I observed that the cells cluster differently depending on some of these variables. Should I regress out these sources of variation (using the SCTransform option vars.to.regress) before performing the differential expression analysis with DESeq2? Or should I perform the differential expression analysis on the raw counts, even though they show these differences?

Another question I have is whether I can use the vars.to.regress option freely with as many variables as I want, or if doing so comes at a cost to accuracy, interpretability, etc.

Answer 1

[fracarb8]

Two things here

• DESeq2, like negbinom an poisson, requires counts, so values not scaled.

• vars.to.regress affects pca, but not DE testing. If you want to regres out some confounders, you should use latent.vars (see FindMarkers docu).

[alvaroperona1]

From what I have been told, the use of FindMarkers overestimates p-values as it treats data coming from the same origin as different samples. After investigating, I found out that you can use the design parameter of DESeq2 so that it takes away the variance that results from other variables by specifying something like: design = ~ variable1 + variable2 + variable3 + variable_of_interest. This way, I can use the normalized raw counts without worrying about variables 1, 2, and 3 masking or altering the effects of the variable of interest. Is this true?